Browse guides and helpSecurity boundaries

Collaborate with informed decisions

Know what you are authorizing.

Understand how files, accounts, and permissions are used before starting. Rivloom currently suits trusted participants and trusted projects.

Rivloom does not provide an operating system security sandbox. AI tools and commands run with the current Windows user's permissions on the execution device. Use trusted projects in trusted environments. Do not accept tasks from strangers or use untrusted projects.

Verify the version you plan to install

The current Rivloom installer is unsigned. Use the download pageto check the build version, file size, and SHA-256. A file hash checks bytes; it is not a publisher signature. Version 0.1.5 and later support in-app updates. Before installation, Rivloom verifies the official update signature, stops its background service, and backs up application data. Back up project files separately.

Choose an approval mode for the task

  • Request approval: Edits, commands, and network access require your approval. Access to directories outside the project is denied.
  • Approve for me: Project edits and commands are approved automatically. Network access still needs approval, and access to directories outside the project is denied.
  • Allow any operation: Within the supported tool set, network access and operations outside the project are also approved automatically.

All three modes restrict direct reads of sensitive files, subagents and upstream native Skill loading. Rivloom shared Skills are read through separate knowledge tools within the granted sharing scope. These rules do not fully isolate arbitrary commands; allowed commands, project configuration and dependencies can still affect other files, access networks or incur costs.

Device trust and operation approval are separate

Discovering a device does not establish trust. Pairing requires confirmation on both devices, and collaboration messages travel over authenticated, encrypted connections. The receiving device manages its own projects, models, account quota, and execution capability. Pairing does not create an enterprise personnel permission system.

Tasks from trusted devices are received automatically. They wait if execution is disabled or resources are not configured, but you should still confirm that you are willing to let the other device use your local projects and model quota before trusting it.

Pausing the queue does not stop running tasks

Pausing and holding affect waiting tasks that have not acquired an execution slot. To end running work, select Stop in the conversation and wait for the result. A communication failure does not prove that the task has stopped.

Stopping does not undo changes already made

Files already written, commands already run, and charges already incurred are not rolled back automatically. Rivloom does not create file snapshots or prevent other editors from changing files at the same time. Keep your own backups and check actual files and test results before accepting a task.

Reading and executing shared content

Registering or reading a Skill does not execute scripts. Shared instructions and memory can influence agent behavior; actual script execution still follows the task approval mode. Use trusted sources. Revoking sharing cannot recall bytes already delivered.

Report security vulnerabilities privately

Report a vulnerability privately on GitHub (sign-in required). Include the affected version, a minimal reproduction and potential impact using synthetic data. Do not disclose exploit details in public issues or pull requests. Only the latest release is maintained; separate backports to older releases are not promised.

For model services, collaboration content, and local records, read the Privacy notice.